The present invention relates to programmable logic devices. Specifically it relates to programmable interconnect that can carry a plurality of input and output signals in a single wire, thus providing a significant reduction of interconnect required for programmable logic devices.
Traditionally, integrated circuit (IC) devices such as custom, semi-custom, or application specific integrated circuit (ASIC) devices have been used in electronic products to reduce cost, enhance performance or meet space constraints. However, the design and fabrication of custom or semi-custom ICs can be time consuming and expensive. The customization involves a lengthy design cycle during the product definition phase and high Non Recurring Engineering (NRE) costs during manufacturing phase. In the event of finding a logic error in the custom or semi-custom IC during final test phase, the design and fabrication cycle has to be repeated. Such lengthy correction cycles further aggravate the time to market and engineering cost. As a result, ASICs serve only specific applications and are custom built for high volume and low cost.
Another type of semi custom device called a Gate Array customizes modular blocks at a reduced NRE cost by synthesizing the design using a software model similar to the ASIC. Structured ASICs provide a larger modular block compared to Gate Arrays, and may or may not provide pre instituted clock networks to simplify the design effort. In both, a software tool has to undergo a tedious iteration between a trial placement and ensuing wire “RC” extraction for timing closure. The missing silicon level design verification in both results in multiple spins and lengthy design iterations, further exacerbating a quick design solution.
In recent years there has been a move away from custom or semi-custom ICs toward field programmable components whose function is determined not when the integrated circuit is fabricated, but by an end user “in the field” prior to use. Off the shelf, generic Programmable Logic Device (PLD) or Field Programmable Gate Array (FPGA) products greatly simplify the design cycle. These products offer user-friendly software to fit custom logic into the device through programmability, and the capability to tweak and optimize designs to improve silicon performance. The flexibility of this programmability is expensive in terms of silicon real estate, but reduces design cycle and upfront NRE cost to the designer.
FPGAs offer the advantages of low non-recurring engineering costs, fast turnaround (designs can be placed and routed on an FPGA in typically a few minutes), and low risk since designs can be easily amended late in the product design cycle. It is only for high volume production runs that there is a cost benefit in using the more traditional approaches. Compared to PLD and FPGA, an ASIC has hard-wired logic connections, identified during the chip design phase. ASIC has no multiple logic choices, no multiple routing choices and no configuration memory to customize logic and routing. This is a large chip area and cost saving for the ASIC. Smaller ASIC die sizes lead to better performance. A full custom ASIC also has customized logic functions which can take less gate counts compared to PLD and FPGA implementations of the same functions. Thus, an ASIC is significantly smaller, faster, cheaper and more reliable than an equivalent gate-count PLD or FPGA. The trade-off is between time-to-market (PLD and FPGA advantage) versus low cost and better reliability (ASIC advantage). The cost of Silicon real estate for programmability provided by the PLD and FPGA compared to ASIC determines the extra cost the user has to bear for customer re-configurability of logic functions.
The ratio of FPGA to ASIC logic gate Silicon area can result as much as 30 to 40 times to implement identical content. Such a large Silicon area disadvantage lead to significant cost and performance disparity between the ASIC and the FPGA. A significant FPGA logic gate Silicon density improvement has been disclosed in the incorporated-by-reference application Ser. Nos. 10/267,483, 10/267,484 and 10/267,511. Such techniques can reduce the ratio of FPGA to ASIC logic gate Silicon area to 3 to 6 times. The most significant portion of Silicon real estate overhead is consumed by programmable interconnects in an FPGA. In a conventional FPGA, over 90% of the configuration memory is dedicated to customize routing for the user, while only under 10% of the configuration memory is utilized to customize logic. Reducing the FPGA logic area penalty to less than 2× would eliminate the need for ASIC designs, and the FPGA design will become the new standard for system design.
In an exact analogy between an FPGA and a City; the Houses in the City are equivalent to Logic in the FPGA, and the Roads in the City are equivalent to Routing Wires in the FPGA. Each Input and Output signal of a Logic Block in the FPGA is a dedicated Incoming and Outgoing Road to the House. Any network of Roads to customize a generic City with specific travel needs would be enormous: first all Houses have to be identified with the correct Incoming & Outgoing Roads, then the required Roads must be found within the network of Roads, and finally the travel times have to be computed to ensure that all meet the Travel Time budget. Not having enough Roads mandate taking detours that affect critical time budgets. Not having enough Red/Green lights and intersections in the network of Road affects the Road utilization efficiency & navigation. The challenge of an FPGA is similar: to provide a network of Roads that do not take up most of the City area, or to keep the FPGA area close to ASIC area.
In a PLD and an FPGA, a complex logic design is broken down to smaller logic blocks and programmed into logic blocks provided in the FPGA. Smaller logic elements allow sequential and combinational logic design implementations. Combinational logic has no memory and outputs reflect a function solely of present inputs. Sequential logic is implemented by inserting memory into the logic path to store past history. Current PLD and FPGA architectures include transistor pairs, NAND or OR gates, multiplexers, look-up-tables (LUTs) and AND-OR structures in a basic logic element. In a PLD the basic logic element is labeled as macro-cell. Hereafter the terminology FPGA will include both FPGAs and PLDs, and the terminology logic element will include both logic elements and macro-cells. Granularity of a FPGA refers to logic content of a basic logic element. Smaller blocks of a complex logic design are customized to fit into FPGA grain. In fine-grain architectures, a small basic logic element is enclosed in a routing matrix and replicated. This is like building one room track Houses in the City. These offer easy logic fitting at the expense of complex routing. In course-grain architectures, many basic logic elements are combined with local routing and wrapped in a routing matrix to form a logic block. This is like building repeating programmable neighborhoods in the City, each neighborhood providing some customization capability. The logic block is then replicated with global routing. Larger logic blocks make the logic fitting difficult and the routing easier. A challenge for FPGA architectures is to provide easy logic fitting (like fine-grain) and maintain easy routing (like course-grain). It's balancing the neighborhood size with the network of roads required.
Inputs and outputs for the Logic Element or Logic Block are selected from the programmable Routing Matrix. A routing wire is dedicated to each. An exemplary routing matrix containing logic elements described in Ref-1 (Seals & Whapshott) is shown in FIG. 1. In that example, the inputs and outputs from Logic Element 101-104 are routed to 22 horizontal and 12 vertical interconnect wires with programmable via connections. These connections may be anti-fuses or pass-gate transistors controlled by SRAM memory elements. These are the Red/Green control lights in the network of Roads comprising a Connect state and a Disconnect state. One output of element 101 is shown coupled to one of the inputs to element 104 in darker lines: in that vertical wire #3 is used to complete the coupling. One output of element 103 is also shown coupled to one of the inputs to element 104 in darker lines: in that vertical wire #8 is used to complete the coupling. Thus every input and every output occupies one or more dedicated wires to complete the coupling. Thus the number wires, wire segments, programmable connection, and Si area required for the connectivity grows rapidly with the number of logic elements N within the fabric.
The logic element having a built in D-flip-flop used with FIG. 1 routing as described in Ref-1 is shown in FIG. 2. In that, elements 201, 202 and 203 are 2:1 MUX's controlled by one input signal each. Element 204 is an OR gate while 205 is a D-Flip-Flop. Without global Preset & Clear signals, eight inputs feed the logic block, and one output leaves the logic block. These 9 wires are shown in FIG. 1 with programmable connectivity. Thus 9 wires must be assigned to connect the logic element shown in FIG. 2. All two-input, most 2-input and some 3-input variable functions are realized in the logic block and latched to the D-Flip-Flop. FPGA architectures for various commercially available devices are discussed in Ref-1 (Seals & Whapshott) as well as Ref-2 (Sharma). A comprehensive thesis on FPGA routing architecture is provides in Ref-3 (Betz, Rose & Marquardt) and Ref-4 (Lemieux & Lewis).
Routing block wire structure defines how logic blocks are connected to each other. Neighboring logic elements have short wire connections, while die opposite corner logic blocks have long wire connections, or a multiple of shorter wires connected to make a long wire. All wires are driven by a fixed pre-designed logic element output buffer and the drive strength does not change on account of wire length. Longer wires may have repeaters to rejuvenate the signals periodically. Buffers consume a large Si area and very expensive. The wire delays become unpredictable as the wire lengths are randomly chosen during the Logic Optimization to best fit the design into a given FPGA. FPGA's also incur lengthy run times during timing driven optimization of partitioned logic. As FPGA's grow bigger in die size, the number of wire segments and wire lengths to connect logic increase. Wire delays dominate chip performance. Wire delays grow proportional to square of the wire length, and inverse distance to neighboring wires. Maximum chip sizes remain constant at mask dimension of about 2 cm per side, while metal wire spacing is reduced with technology scaling. A good timing optimization requires in depth knowledge of the specific FPGA fitter, the length of wires segments, and relevant process parameters; a skill not found within the design house doing the fitting. In segmented wire architectures, expensive fixed buffers are provided to drive global signals on selected lines. These buffers are too few as they are too expensive, and only offer unidirectional data flow. Predictable timing is another challenge for FPGA's. This would enhance place and route tool capability in FPGA's to better fit and optimize timing critical logic designs. More wires exacerbate the problem, while fewer wires keep the problem tractable, reducing FPGA cost.
FPGA architectures are discussed in detail in the referenced US patents incorporated herein by reference. These patents disclose specialized routing blocks to connect logic elements in FPGA's and macro-cells in PLD's. In all cases a fixed routing block is programmed to define inputs and outputs for the logic blocks, while the logic block performs a specific logic function. Such dedicated interconnect wires drive the cost of FPGAs over equivalent functionality ASICs.
Four methods of programmable point to point connections, synonymous with programmable switches, between A and B are shown in FIG. 3. These are the equivalence of Red/Green signal lights in FPGAs. A configuration circuit (the method to decide and change Red vs. Green) to program the connection is not shown. All the patents listed above use one or more of these basic connections to configure logic elements and programmable interconnect. The user implements the decision by programming memory. In FIG. 3A, a conductive fuse link 310 connects A to B. It is normally connected, and passage of a high current or a laser beam will blow the conductor open. In FIG. 3B, a capacitive anti-fuse element 320 disconnects A to B. It is normally open, and passage of a high current will pop the insulator to short the terminals. Fuse and anti-fuse are both one time programmable due to the non-reversible nature of the change. In FIG. 3C, a pass-gate device 330 connects A to B. The gate signal S0 determines the nature of the connection, on or off. This is a non destructive change. The gate signal is generated by manipulating logic signals, or by configuration circuits that include memory. The choice of memory varies from user to user. In FIG. 3D, a floating-pass-gate device 340 connects A to B. Control gate signal S0 couples a portion of that to floating gate. Electrons trapped in the floating gate determines on or off state of the connection. Hot-electrons and Fowler-Nordheim tunneling are two mechanisms to inject charge onto floating-gates. When high quality insulators encapsulate the floating gate, trapped charge stays for over 10 years. These provide non-volatile memory. EPROM, EEPROM and Flash memory employ floating-gates and are non-volatile. Anti-fuse and SRAM based architectures are widely used in commercial FPGA's, while EPROM, EEPROM, anti-fuse and fuse links are widely used in commercial PLD's. Volatile SRAM memory needs no high programming voltages, is freely available in every logic process, is compatible with standard CMOS SRAM memory, lends to process and voltage scaling and has become the de-facto choice for modern very large FPGA devices.
A volatile six transistor SRAM based configuration circuit is shown in FIG. 4A. The SRAM memory element can be any one of 6-transistor, 5-transistor, full CMOS, R-load or TFT PMOS load based cells to name a few. Two inverters 403 and 404 connected back to back forms the memory element. This memory element is a latch. The latch can be full CMOS, R-load, PMOS load or any other. Power and ground terminals for the inverters are not shown in FIG. 4A. Access NMOS transistors 401 and 402, and access wires GA, GB, BL and BS provide the means to configure the memory element. Applying zero and one on BL and BS respectively, and raising GA and GB high enables writing zero into device 401 and one into device 402. The output S0 delivers a logic one. Applying one and zero on BL and BS respectively, and raising GA and GB high enables writing one into device 401 and zero into device 402. The output S0 delivers a logic zero. The SRAM construction may allow applying only a zero signal at BL or BS to write data into the latch. The SRAM cell may have only one access transistor 401 or 402. The SRAM latch will hold the data state as long as power is on. When the power is turned off, the SRAM bit needs to be restored to its previous state from an outside permanent memory. In the literature for programmable logic, this second non-volatile memory is also called configuration memory. The SRAM configuration circuit in FIG. 4A controlling logic pass-gate as shown in FIG. 3C is illustrated in FIG. 4Ba. Element 450 represents the configuration circuit. The S0 output directly driven by the memory element in FIG. 4A drives the pass-gate gate electrode. In addition to S0 output and the latch, power, ground, data in and write enable signals in 450 constitutes the SRAM configuration circuit. Write enable circuitry includes GA, GB, BL, BS signals shown in FIG. 4A. The symbol used for the programmable switch comprising the SRAM device and the pass-gate is shown in FIG. 4Bb as the cross-hatched circle 460. SRAM memory data can be changed anytime in the operation of the device, altering an application and routing on the fly, thus giving rise to the concept of reconfigurable computing in FPFA devices.
A programmable MUX utilizes a plurality of point to point switches. FIG. 5 shows three different MUX based programmable logic constructions. FIG. 5A shows a programmable 2:1 MUX. In the MUX, two pass-gates 511 and 512 allow two inputs I0 and I1 to be connected to output O. A configuration circuit 550 having two complementary output control signals S0 and S0′ provides the programmability. When S0=1, S0′=0; I0 is coupled to O. When S0=0, S0′=1; I1 is coupled to O. With one memory element inside 550, one input is always coupled to the output. If two bits were provided inside 550, two mutually exclusive outputs S0 and S1 could be generated. That would allow neither I0 nor I1 to be coupled to O, if such a requirement exists in the logic design. FIG. 5B shows a programmable 4:1 MUX controlled by 2 memory elements. A similar construction when the 4 inputs I0 to I3 are replaced by 4 memory element outputs S0 to S3, and the pass-gates are controlled by two inputs I0 & I1 is called a 4-input look up table (LUT). The 4:1 MUX in FIG. 5B operate with two memory elements 561 and 562 contained in the configuration circuit 560 (not shown). Similar to FIG. 5A, one of I0, I1, I2 or I3 is connected to O depending on the S0 and S1 states. For example, when S0=1, S1=1, I0 is coupled to O. Similarly, when S0=0 and S1=0, I3 is coupled to O. A 3 bit programmable 3:1 MUX is shown in FIG. 5C. Point D can be connected to A, B or C via pass-gates 531, 533 or 532 respectively. Memory elements 571, 572 and 573 contained in a configuration circuit 570 (not shown) control these pass-gate input signals. Three memory elements are required to connect D to just one, any two or all three points. In reconfigurable computing, data in memory elements 571, 572 and 573 can be changed on the fly to alter connectivity between A, B, C and D as desired.
In FPGA's the configuration memory content is very high. It is typically loaded when the device is powered up, and takes up a considerable time to fully load all the data. On the fly alteration of memory is extremely cumbersome. References U.S. Pat. Nos. 5,629,637 and 6,480,954 disclose some methods to make the task manageable. The problem has now grown three fold: (i) Dedicated wires were needed to connect the inputs and outputs leading to a large chip area, (ii) Extra circuitry is inserted to identify which portion of the memory data is refreshed further adding to the cost of the device, and (iii) The operation must be halted to upgrade the memory with new data leading to significant switch-over dead time. Efficient software tools that can synthesize designs into a multitude of variable designs do not exist even in these modern days. Reconfigurable computing does not resolve the high cost of FPGAs over ASICs.
What is desirable is to reduce the Silicon overhead required to support routing wires within a programmable logic device. The routing must provide timing predictability and easily integrated into a software tool. These routing connections need to facilitate short wire connections and long wire connections and then preserve timing in a predictable and calculable manner. One method to reduce the wire overhead is to provide a programmable time multiplexing scheme to share one wire with a plurality of inputs/outputs, thus reducing the overall wires needed within the FPGA. It is also beneficial to have the ability to program the data flow direction, and have the entire configurability integrated into vertical configuration circuits. Vertically integrated configuration circuits in 3D FPGAs, previously presented and incorporated herein, provide significant cost reductions and performance improvement to FPGAs. Previously presented techniques incorporated herein to use bi-directional buffers and highly efficient bridge structures all consume less Si real estate to further reduce the cost of FPGAs. The new interconnect structure must reach reasonable cost parity to ASICs (within 2× of ASIC cost) and also lend to an easy application specific design conversion to the user, preserving the original timing characteristics of the circuit during the conversion.